Abstract: For rapid determination of tar, nicotine and CO in mainstream smoke released from tobacco leaves stored in warehouses of different tobacco enterprises, near-infrared(NIR)spectra of 889 tobacco strip samples were collected and the Kennard-Stone(K-S)algorithm was used to select 150 representative tobacco strip samples for Partial Least Squares (PLS) regression modeling. The spectra were pre-processed using Multivariate Scatter Correction (MSC), Savitzky-Golay (SG) smoothing and first order derivatives. By selecting the characteristic spectral segments and the optimal number of principal components, the prediction models for the releases of tar, nicotine and CO in cigarette smoke based on the NIR spectroscopy of tobacco leaves were established. The results showed that: 1)There were no significant differences between the results of the prediction models and the standard methods at 0.05 level of significance. 2)All the coefficients of determination(R²)of the best PLS prediction models for tar, nicotine and CO releases were higher than 0.80. 3)The root mean square errors of calibration(RMSEC) for tar, nicotine and CO were 1.13, 0.13 and 0.97, respectively. The ratios of the root mean square error of prediction(RMSEP)to the RMSEC for tar, nicotine and CO were 1.03, 1.00 and 0.95, respectively. 4)The results of the leaf blending overlay experiment showed that the predicted results of each tobacco raw material linearly added according to the formula ratios, which closely matched the actual determined result of the leaf blending with a relative deviation of less than 10%. The established near-infrared prediction models for the releases of tar, nicotine and CO were therefore considered to be accurate, reliable and suitable for rapid quantitative prediction of the releases of routine cigarette smoke components from tobacco leaves.